Duboko učenje korištenjem YOLO modela za nedestruktivna ispitivanja

master's thesis

• Title (english): Deep Learning Using YOLO Model for Nondestructive Testing
• Author: Luka Posilović
• Mentor: Sven Lončarić (mentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Tomislav Petković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2019-07-10, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering

Abstract

Nedestruktivno ispitivanje je metoda kojom se pronalaze pukotine i ostale nesavršenosti u materijalima bez njihova uništavanja. Vrlo je važna prilikom ispitivanja konstrukcijskih elemenata u nuklearnim elektranama, ali i mnogim drugim industrijama poput avio industrije i automobilske industrije. Jedna od metoda nedestruktivnog ispitivanja koristi se ultrazvučnim valovima kako bi pronašla pukotine i odredila njihov tip i veličinu. Za potrebe ispitivanje nekog objekta snimi se velika količina ultrazvučnih slika koje ljudski ekspert mora detaljno pregledati. Mogućnost ljudske pogreške i vrijeme potrebno za kvalitetnu analizu svih podataka otežavajući su faktori u provedbi ispitivanja. Ovom problemu pristupa se razvojem algoritma za automatsku detekciju defekata u ispitivanom materijalu. Razvijeni algoritam je konvolucijska neuronska mreža koja na ulazu prima ultrazvučne slike, a izlaz su je pozicija i veličina pronađenih defekata. Korišten je model YOLO - You Only Look Once s pretreniranim baznim slojevima. Model je treniran koristeći bazu od 490 slika podijeljenih na skup za treniranje, validaciju i testiranje. Predstavljeni model ostvario je prosječnu točnost (mAP) od 93,5% na testnom skupu podataka s mogućnošću obrade 45 slika u sekundi. Također, razvijen je algoritam za uklanjanje šuma na ultrazvučnim slikama koristeći valićnu transformaciju.

Abstract (english)

Non-destructive testing is a method for detecting flaws and other impurities in materials without destroying them. It is very important during inspection of construction blocks in nuclear power plants, but also many other different industries such as aviation industry and automotive industry. One of the methods of non-destructive testing is using ultrasonic waves to find cracks and determine their shape and size. To inspect some object a vast amount of ultrasonic images is acquired that a human expert needs to inspect in detail. Possibility of human error and time needed for a thorough inspection of all data are aggravating factors in inspection and analysis. The approach to this problem is the development of an algorithm for automatic flaw detection in inspected materials. Developed algorithm is a convolutional neural network which inputs ultrasonic images and outputs the position and size of detected defects. Model is trained using the database of 490 images divided into the set for training, validation and test. Presented model achieved the average precision (mAP) of 93,5% on a test set while analyzing 45 images per second. Also, an algorithm for denoising ultrasonic images using wavelet transform is presented.

• Keywords: obrada slike; analiza slike; ultrazvucno snimanje; nedestruktivno testiranje; konvolucijske neuronske mreže; automatska inspekcija
• Keywords (english): image processing; image analysis; ultrasonic scanning; nondestructive testing; convolutional neural network; automated inspection
• Language: croatian
• URN:NBN: urn:nbn:hr:168:152395
• Study programme: Title: Electrical Engineering and Information Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije (magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2020-01-07 21:29:48